Project 4: Methods for Discovery and Evaluation of Dynamic Treatment Regimes PROJECT SUMMARY Treatment of cancer is an ongoing process during which clinicians make a series of therapeutic decisions over the course of the disease. The prospect of using a patient's genomic information, derived biomarkers, and other baseline and evolving characteristics to guide this sequential decision-making process in an evidence-based way would be a significant step toward personalized cancer medicine. Accordingly, there has been increasing interest in identifying overall strategies for sequential decisions that lead to the most beneficial outcomes, where the treatment selected for a patient at each decision is predicated using a synthesis of potentially complex accruing patient information. However, most current cancer clinical trials evaluate only the therapeutic options available at a single decision point. Combining results across distinct single decision point studies to deduce an overall strategy cannot take into account the fact that the selection of treatment at a particular decision point must be placed in the context of prior and subsequent decisions. For example, a treatment may have prolonged effects that impact the efficacy of future treatments. Moreover, the therapies that may be beneficial over the course of a disease for one patient may differ from those for another. The entire sequential decision process must be studied as a whole. A principled framework for this object is to consider cancer treatment strategies as dynamic treatment regimes, which are formal algorithms for sequential decision-making that use information on the patient accrued at each decision point to determine the next step of treatment. Methods for discovering the optimal regime (the regime that leads to the most beneficial outcomes if used by the patient population), have the potential to transform how cancer treatment is viewed. The four specific aims of this project seek to catalyze this paradigm shift by addressing specific challenges that arise in the cancer context. Because balancing competing outcomes such as efficacy and toxicity is a key goal of cancer therapy, the first aim is focused on methods for the discovery of optimal regimes that target this balance. The second aim proposes new methods for the design of cancer clinical trials for which the primary goal is to estimate an optimal regime. Outcomes of primary interest in cancer are often censored times-to-an-event (e.g., survival), but there is little work on estimating optimal regimes based on maximizing survival probabilities and related quantities relevant in the cancer context; new methods are proposed in the third aim. An optimal regime based on distilling high-dimensional information, e.g., genomic data, may be too complex to implement or interpret. The fourth aim continues our work on methods for discovery of optimal regimes restricted to be in a class of practically- feasible, interpretable regimes. Collectively, these aims will yield a suite of methodological advances that will enable cancer researchers to develop personalized, evidence-based, sequential decision-making strategies. The benefit to cancer patients will be substantial as the methods will allow the best possible treatment decisions to be made for individual patients over the course of their disease based on their own information.